Posted: 02-06-2024

Location: U.S. Army Combat Capabilities Development Command – Aberdeen Proving Ground

Level: Graduate Student 

General Topic: Computer Science

Description of Research:Most Army systems have limited computation, communication, and power capabilities due to requirements for low Size, Weight, and Power (SWaP). Tactical devices may not have sufficient computational capability to process the live video, audio, or images from sensors to meet real-time mission requirements. As the Army is embracing AI to accomplish its mission goals, resource constraints of tactical computing platforms will impede the deployment of complex AI algorithms over tactical platforms. AI model optimization algorithms will be researched to reduce the model complexity and accelerate inference to achieve real-time performance in resource-limited edge platforms. Tactical unmanned ground vehicles (UGVs) with limited computing resources are reliant upon resource constraint-aware adaptive computing algorithms for real-time operations. Autonomous AI stacks in UGVs utilize complex deep neural network (DNN) algorithms for intelligent navigation, computer vision and for exhibiting tactical behaviors in contested and constrained environments. Only after computational complexity reduction and model optimization of DNN algorithms are these algorithms executable for real-time on UGVs with limited resources. The aim of the project is to develop MDO-capable optimized, adaptive algorithms for reducing computational complexity of various learning, perception algorithms associated with the Autonomy AI stack supporting autonomy on UGVs. There are three technique approaches to optimize the DNN models in our efforts. First is to optimize the perception algorithm by reducing the computational complexity. Second is to optimize the neural networks to accelerate inference by exploiting the hardware properties. Third is to develop adaptive algorithms that could adapt to the device resource constraints to allow algorithms to meet their performance objectives. All these approaches are complementary to each other and can be combined to achieve better performance.